GLM-5.3's 14-day fuse: a 2,436-bug discovery model heads for open weights

Zhipu put GLM-5.3 behind its GLM Coding Plan on August 14, 2026, and said the weights follow in roughly two weeks, once security reviews finish ^[1][2]. That interval is the story: for about fourteen days, a model the company credits with 2,436 vulnerabilities across 269 projects exists only as a service somebody can throttle ^[8][20].

What the figure turns on is which half of the offensive workflow travelled. On CyberGym, which tests whether a model can identify and validate flaws from source code, Zhipu reports 84.5 per cent, ahead of Anthropic's Mythos 5 at 83.8 and OpenAI's GPT-5.6 Sol at 83.6 ^[4]. On ExploitBench, which measures how far a model climbs the exploitation ladder, it reports 54.4 per cent against 78 and 76.5, leaving it 23.6 points behind Mythos ^[5][15]. So the capability about to become un-gatable is the finding, not the finishing. The trajectory is the part worth noting: the ExploitBench score more than doubled from GLM-5.2's 24.4 per cent, and on ExploitGym the model went from 29 completed tasks in two hours to 105, on the same base model with scaled post-training and added vulnerability-discovery data ^[6][7][12].

The ledger is where the two weeks bite. Zhipu lists 107 critical and 990 high-severity findings, which is exactly its 1,097 medium-to-high total; 53 are published and 2,383 remain under embargo, spanning kernels, operating systems, browser engines, open-source infrastructure and network protocols ^[9][18][17]. The company has not said how many were previously unknown, or how many anyone independently reproduced ^[11]. All of the scores are Zhipu's own evaluations, with no independently reproduced benchmark set available yet ^[3]. Neil Shah of Counterpoint Research put the constraint plainly: once weights are released freely, built-in guardrails can be stripped without any cognizance or control ^[13].

For scale on the defender side, CISA's implementation guidance for BOD 26-04 allots agencies the first two hours after a KEV addition to decide whether one CVE meets the under-three-days remediation threshold ^[16].

Watch the license and model card when the files publish, and whether the 2,383 embargoed entries begin moving through disclosure once anyone can rerun the discovery pipeline locally ^[9][2].